Function create_padding_mask 

pub fn create_padding_mask(
    ctx: &WgpuContext,
    lengths: &[usize],
    max_seq_len: usize,
) -> UnifiedTensor

Helper to create a padding mask for Flash Attention V2 (additive mask)

Returns a tensor of shape [batch_size, 1, 1, max_seq_len] (or broadcastable).

• 0.0 for valid positions
• -1e9 for padded positions